Researchers from Skoltech and KU Leuven Employ Deep Learning to Augment 3D Micro-CT Images of Fibrous Materials Using Neural Networks

Researchers from KU Leuven and Skoltech utilized machine learning to help recreate three-dimensional micro-CT images of fibrous materials. This task, which is essential for sophisticated material analysis, is complex and time-consuming for humans. This research work was published in Computational Materials Science Journal.

Micro-computed tomography is incredibly helpful for studying the 3D microstructure of fiber-reinforced composites and other complex materials. It is, however, a picky tool: samples are small, and photos frequently contain abnormalities such as darkened, missing, or damaged regions. Researchers took inspiration and knowledge from the art industry to help them deal with this, where damaged artworks must be restored while maintaining their overall integrity. As a result, inpainting has become a standard digital image processing technique.

According to Radmir Karamov, the paper’s first author and a PhD student at Skoltech and KU Leuven, the key advantage of AI in painting is speed. They can process a hundred photos per second using a trained model, which would take a human much longer. Also, computers are far more superior at working with a three-dimensional (3D) image because they see it from all sides as well as right through and can instantaneously reconstruct the entire volume and not just the surface as we human beings do.

To cover a gap in the range of current inpainting methods for 3D micro-CT images, the team used 3D encoder-decoder generative adversarial networks or GANs.

According to scientists, reinforcing inclusions in composite materials, such as fibers can be randomly orientated in three dimensions. As a result, scientists must use 3D images to describe this intricate interior substructure. The scientists went to GANs since more traditional convolutional neural networks couldn’t achieve the precision required for this assignment.

Karamov explained how researchers train two competing networks in GANs rather than train a single neural network to reconstruct pictures. A generator network attempts to create realistic-looking fake images, while a discriminator network evaluates the images and assesses whether they are genuine or not. You can think of this as a battle between counterfeiters and the police, according to Goodfellow, the GANs developer. Counterfeiters want to make and produce fake money that looks real, and the police want to find out whether it is fake or not by looking at any particular bill.

The researchers investigated three GAN topologies on micro-CT scans of short glass fiber composites with no repeating structures, the most challenging scenario for inpainting. They chose the architecture that offered good inpainting quality and performance with minimal GPU memory utilization.

They can also use the inpainting algorithm to remove all defects in micro-CT scans for more precise modeling of material behavior and examine how material performance would improve if all internal voids and pores are removed during the manufacturing process, added Karamov.

According to the researcher, inpainting is merely the first step in developing a completely automated generative algorithm for novel materials, which would allow scientists to build a material based on the qualities required for a specific application.